yangguoquan001

Results 6 comments of yangguoquan001

> [@mihongyu](https://github.com/mihongyu) edit_image 是编辑模型的输入图像路径,也就是编辑前的图像路径。 > > 如果删除这一列(包括csv文件的表头)并改为--data_file_keys "image",那么训练框架将会把这一模型当作标准文生图模型进行训练。 我靠!原来是这样,我一直搞反了

> > [@mihongyu](https://github.com/mihongyu) edit_image 是编辑模型的输入图像路径,也就是编辑前的图像路径。 > > 如果删除这一列(包括csv文件的表头)并改为--data_file_keys "image",那么训练框架将会把这一模型当作标准文生图模型进行训练。 > > [@Artiprocher](https://github.com/Artiprocher) 这个命名很容易让人confuse,以为edit_image才是编辑后的图片,建议改成叫cond_image比较好 我才发现我就搞错了,艹

> Found that I could get 2509 running by ticking `cache_text_embeddings: true` when encountering this error worked for me!

--lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1" \ --lora_rank 32 \ --use_gradient_checkpointing \ --dataset_num_workers 8 \ --find_unused_parameters \ --enable_fp8_training 我不适用deepspeed在48G显存上按照这个参数是可以训练的

I also encounter the same problem: the loss intend to oscillation instead of converging when fine tuning qwen-image-edit and qwen-image-edit-2509 with lora.

https://github.com/ostris/ai-toolkit/issues/472 I solved this problem according to the comment of this issue.